Learning to Manipulate Object Collections Using Grounded State Representations

We propose a method for sim-to-real robot learning which exploits simulator state information in a way that scales to many objects. We first train a pair of encoder networks to capture multi-object state information in a latent space. One of these encoders is a CNN, which enables our system to operate on RGB images in the real world; the other is a graph neural network (GNN) state encoder, which directly consumes a set of raw object poses and enables more accurate reward calculation and value estimation. Once trained, we use these encoders in a reinforcement learning algorithm to train image-based policies that can manipulate many objects. We evaluate our method on the task of pushing a collection of objects to desired tabletop regions. Compared to methods which rely only on images or use fixed-length state encodings, our method achieves higher success rates, performs well in the real world without fine tuning, and generalizes to different numbers and types of objects not seen during training.